Google hasn't released details about their model yet, but the VGG team in close 2nd place had a very simple and beautiful "vanilla" CovnNet model, but with more

Google hasn't released details about their model yet, but the VGG team in close 2nd place had a very simple and beautiful "vanilla" CovnNet model, but with more careful hyperparameter settings and training/testing protocol. In other words, the source of improvement are tweaks on the original Krizhevsky architecture from 2012, not completely new and unexpected ideas. That is not to belittle the contribution - these experiments take forever to run and require very good practices and intuitions for what to try next. Karen released the details of the model only yesterday on arXiv (http://arxiv.org/pdf/1409.1556v1.pdf http://arxiv.org/pdf/1409.1556v1.pdf) More specifically, as can be seen in the paper, it seems that very deep stacks of conv/conv/pool modules with tiny 3x3 filters work well (which is satisfying because it's really simple and beautiful), and from being more thorough with the training and testing protocol (data augmentations, averaging, multiscale approaches in both train/test time, etc). Google will release details of their method on Sept 12 so we'll know more. From their abstract, it seems they have a more significant departure from a basic convnet architecture.